acceptodds
Under review as a conference paper at ICLR 2027

To Untangle Multi-Circuit Neural Dynamics, All You Need Is Lag

Abstract

The brain computes through interactions between neural populations that shape how their activity evolves over time. These interactions often span multiple brain regions, thus forming brain-wide networks that reconfigure as the animal shifts task contexts, behaviors, and decisions. Within these networks, neural signals travel at finite speed, so one population drives another only after a delay whose length reflects both the wiring involved and the animal's current state. Dynamical systems are a natural way to model such lagged interactions, yet existing models often assume a fixed lag, allow several lags but fix the interaction structure in time, or permit richer lag structure at the cost of uninterpretable dynamics. Here, we present Multi-Lag decomposed Linear Dynamical Systems (ML-dLDS), a framework that captures how time-varying neural dynamics arise from multiple co-occurring processes acting across several lags. ML-dLDS represents neural activity as a sparse decomposition of core neural interactions, or circuits, which are reused over time and act across a sparse set of lags. Each circuit-lag pair carries its own time-varying coefficient that sets its momentary contribution to the overall network dynamics, which thereby shift from moment to moment as the active circuit combination changes. We fit the circuits, their per-lag gains, and their time-varying coefficients through an alternating optimization under structured-sparsity penalties. We first validate ML-dLDS on synthetic datasets, where it recovers the ground-truth circuits and their lags, and then demonstrate it on multi-region electrophysiological recordings, where it reveals lag structure that single-lag models oversimplify. Finally, we apply ML-dLDS to COVID-19 case data, where it uncovers interpretable multi-lag structure across countries.

Primary area:
applications to neuroscience & cognitive science
Keywords:
neural population dynamics, dynamical systems, interpretability, time-varying dynamics, multiple lags, delayed interactions, structured sparsity

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Dfp8OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes ML-DLDS that learns interpretable dynamics across multiple lags. ML-DLDS models interaction via linear mappings with lag gains, time-varying coefficients. The model is optimized by assuming structured sparsity on lag gains SS and coefficients cc. The proposed method is evaluated on synthetic datasets, Neuropixels recordings and COVID time series. Overall, I think the neuro-science motivation is good, but the contribution appears to be an incremental extension of decomposed linear dynamical systems with limited technical novelty and insights. Experimental results are not very convincing in that it does not show meaningful structures recovered and why/when ML-DLDS is better than the existing methods.

Critical strengths

The neuro-science background of the paper is appreciated: I agree that understanding how multiple neural populations interact across different timescales is an important problem. The authors have done a good job in positioning the paper among the literature. Technically, sharing interaction matrices across delays is a sensible modeling choice that could provide a interpretable description of these interactions. Experimental results show that to some extent incorporating multiple delays is useful.

Critical weaknesses

I am afraid the paper has limited novelty overall. ML-dLDS is based on dLDS, which already represents dynamics through sparse combinations of shared interaction matrices. Compared against existing methods, the main contribution of the paper seems a sparse lag profile for simultaneously active shared circuits. However, the paper provides little technical insight into what advantage ML-dLDS offers over the existing methods. In the related work section, these methods are discussed but not evaluated experimentally. I also find that the synthetic datasets seem to match the assumptions. For hard problems, these assumptions do not necessarily hold.

Questions for authors
  • what is the main advantage and improvement over existing methods like dLDS besides allowing multiple lags? Does ML-dLDS provide new insights beyond the baselines?
  • is it possible to compare against stronger methods that already model multiple delays or nonlinear interactions?
  • do the neural findings hold across additional recording sessions or animals? I notice that only one session is used in the Neuropixels dataset. How representative is this?
2 comments
AuthorsOpenReview ↗

Rebuttal, part 1, Dfp8

We thank the reviewer for the careful reading and for recognizing the value of modeling multi-lag dynamics. We address each point below.

W1 and Q1 (part a). Main advantage and improvement over existing methods.

We thank the reviewer and agree that dLDS is the foundation of our model. We claim, however, that ML-dLDS is not a straightforward extension of dLDS to multiple lags. In particular, a naive multi-lag extension of dLDS would suffer from the natural ambiguity of the system, as its many parameters admit multiple equivalent solutions when past states at different lags carry similar information. We kindly refer the reviewer to the original submission, where we discussed this inherent ambiguity (Sec. 3.2, App. J). This challenge requires a new model that copes with this ambiguity via structured regularization, and accordingly also requires a modified inference procedure. The novelty of ML-dLDS is therefore fourfold: 1) a lag matrix S\textbf{S} with structured regularization, 2) a modified inference procedure, 3) a new initialization, and 4) recovery of the lags at which each circuit acts, which a standard dLDS would obscure and simplify into a single lag.

For the first, the lag matrix S\textbf{S} assigns each circuit a sparse set of lags, and its structured penalty selects among the equivalent solutions (Eq. 4). For the second, the inference accounts for the coupling of each circuit across its lags when it updates the coefficients, the lag gains, and the circuits (Sec. 3.3). For the third, the initialization detects segments dominated by a single circuit and reads the lag of each circuit from them (App. A). For the fourth, we demonstrated the advantage empirically on multiple experiments. For example, on the mouse recordings ML-dLDS finds interactions from the lateral posterior thalamus at δ=1\delta = 1 that re-emerge at δ=4\delta = 4 (Fig. 4D), and on the COVID-19 data short delays dominate the synchronized waves while longer delays emerge in the desynchronized phase (Fig. 5E).

We also kindly refer the reviewer to the experiments in our paper where we actively tested this effect. In particular, when we removed the lag penalty on synthetic data, the lag-support recovery dropped sharply (Fig. 3G, Table 2). In addition, single-lag dLDS reached a much lower circuit correlation than ML-dLDS (Table 1). We believe that these changes are enough to make ML-dLDS a standalone model.

W2 and Q1 (part b). Technical insight into the advantage.

We appreciate the reviewer's recognition that more clarification and technical insight into the advantage of ML-dLDS would improve the paper. Following the reviewer's concern we clarify the mechanism behind this advantage. When the true dynamics act at lags larger than one, a model limited to a single lag would absorb the earlier states into its operator. For example, if the true dynamics are xt=fxt−2x_t = f x_{t-2}, the best single-lag fit xtx_t ~ Mxt−1M x_{t-1} gives M=fC21C11−1M = f C_{21} C_{11}^{-1} (for simplicity, assuming centered, stationary data), where C21C_{21} is the covariance between xt−2x_{t-2} and xt−1x_{t-1} and C11C_{11} is the covariance of xt−1x_{t-1}. The recovered operator therefore mixes the true circuit ff with the state's own autocorrelation. These mixed operators will differ from the true circuit and will not allow us to know at which lag the circuit acts. In contrast ML-dLDS explicitly learns the lag at which each circuit operates and recovers circuits that are closer to the true ones on synthetic data (Table 1, App. J). With respect to VAR specifically, VAR does model multiple lags but learns one fixed matrix per lag. That means that it cannot capture multiple co-occurring processes at a time and cannot reuse the same operator across multiple timescales. Consequently, ML-dLDS can capture richer dynamics with a comparable number of unknown parameters.

AuthorsOpenReview ↗

Rebuttal, part 2, Dfp8

W3 and Q2. Comparison to related methods.

We thank the reviewer for emphasizing the importance of strong baselines, and we fully agree that there are always further methods one could compare to, including all those we mention in the related work. In the original submission we compared ML-dLDS against four to six baselines across its experiments (App. C, Table 2, Table 3). When designing the experiments, we focused on baselines that produce components equivalent to those of ML-dLDS, such that the comparison can show whether ML-dLDS recovers the underlying circuits, dynamics, and lags less/more accurately than they do.

Regarding the reviewer’s question on the possibility of comparing against methods that model multiple delays or nonlinear interactions, these methods do not provide components analogous to ours, which makes a comparison ill-defined. For example, a comparison to LFADS (Pandarinath et al., 2018) could be made on data fidelity, but it would not allow a proper comparison of the latent dynamics / circuits, as LFADS does not learn a set of circuits analogous to ours (Sec. 1).

Multi-lag methods like mDLAG (Gokcen et al., 2023) would not address the comparison of the dynamical process we focus on either, as mDLAG is not tailored to learn the dynamical operator that explains the temporal evolution. mDLAG instead models each brain area as a combination of a shared set of Gaussian-process latents that appear with a different delay in each area. That means that also the definition of a lag in mDLAG differs from our definition, as a lag in mDLAG is the time offset at which a shared latent appears in each area, whereas in our model a lag is how far back in time a circuit (i.e., the dynamical operator that governs the latent dynamics) reaches to drive the present state (Sec. 2). In addition, the delay in mDLAG is a single fixed value per latent and per area that stays constant over the recording, whereas in our model the lags at which each circuit acts vary in time through the coefficients. The two lags thereby capture different quantities, and only the cross-area part of our lags has a counterpart in mDLAG, as mDLAG has no notion of a circuit that acts within an area or across several lags. We therefore do not necessarily expect to observe an agreement between them.

If the reviewer has a specific method in mind that they believe produces analogous components and should have been included, we would be happy to add it to the revised paper.

W4. Synthetic data and the model assumptions.

We thank the reviewer for noting that the synthetic datasets match the model's assumptions, and we agree this is an important test. We therefore added three new experiments, each across five seeds, 1) synthetic data in which every circuit acts at the same single lag, 2) data generated from a vector autoregression with a separate matrix per lag, and 3) data generated from nonlinear delayed dynamics.

For the first, we tested whether ML-dLDS recognizes the single lag or instead spreads the dynamics across longer lags, and found that it kept lag one as the dominant lag of every circuit (Fig. 15). For the second and third, we tested whether ML-dLDS recovers the correct lags on VAR-driven and on nonlinear data, and found that it recovered the dominant lag of every circuit on both, with a lag-profile correlation of 0.9760±0.0100.9760 ± 0.010 on the VAR data and 0.9758±0.0080.9758 ± 0.008 on the nonlinear data (Fig. 17).

Q3. Additional sessions.

We agree that an additional neural data session could further strengthen our paper. The data are large and require compute access that is hard to get within a short rebuttal period. We are working on it and will include the additional session in the revised paper.

We hope these clarifications and the new experiments address the reviewer’s concerns and questions.

Reviewer_p4svOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Multi-Lag decomposed Linear Dynamical Systems, an architecture that fits neural activity (or other dynamical systems) as a generalized VAR model that decomposes into separate circuits, each of which operates at a sparse set of lag times. Using synthetic data where nontrivial time-delays in interactions are present, they show that their model and fitting method recover the underlying latent dynamics and lag coefficients well, whereas other LDS variants cannot. They then apply their method to neural recordings as well as COVID-19 case data, and argue that interpretable lag structure consistent with experimental observations is recovered.

Critical strengths
  • The central question of how how to capture interactions in systems with heterogenous delays is well-motivated, and the writing is clear and organized overall.
  • The proposed method is tested across a good mix of synthetic datasets with known ground truth and real datasets, and seems to work in these settings with far fewer parameters than standard VAR models.
Critical weaknesses
  • The model introduced has many moving parts that are all jointly fitted, which make me concerned about the robustness of the fitting procedure. In particular, the activation coefficients cktδc_{kt}^{\delta} have no underlying dynamics model to constrain them, and seem free to vary discontinuously both across time tt and across lag value δ\delta. This choice would seem to introduce a ton of solution degeneracy, in contrast to what is asserted in line 176 "...they resolve the ambiguity of the fit, where many activation patterns would otherwise reproduce the same xt\mathbf{x}_t". Or is it actually the case that this fitting procedure estimates similar lag patterns across initialization seeds/reasonable choices of hyperparams?

  • If I understand correctly, the synthetic experiments were generated from a particular instantiation of this very model as ground truth, so it is unsurprising that only this model and not the AR(1)-like LDS variants can capture the interesting lag structure present. Following up on the previous comment: how about the opposite scenario, where the ground truth has no interesting lag structure; under the same hyperparam settings, does this model avoid spuriously attributing dynamics to lags beyond 1?

I would be open to raising my score if these concerns can be addressed.

Questions for authors

See above.

1 comment
AuthorsOpenReview ↗

Rebuttal, p4sv

We thank the reviewer for the constructive feedback on robustness and on testing lag recovery beyond the model. We address each point below.

W1. Robustness of the fit and stability of the recovered lags across seeds and hyperparameters.

We thank the reviewer for suggesting additional priors over the coefficients and for raising the need to check their robustness. We confirm that the coefficients have no explicit underlying dynamical model and we want to clarify that this was an intentional choice. This choice enables the circuits to capture both slowly varying changes and more abrupt switches such that the model can be seen as an extension of switching models with switching models being a special case. We agree that this choice may make the coefficients potentially less robust than with a dynamical model and we address this in our case by applying an ℓ1\ell_1 regularization that keeps only a few circuits active at a time and thereby stabilizes the coefficients. In addition, we note that within the SPGL1 solver we use we set the initial search value of each coefficient at time tt to the coefficient at the previous time point, which introduces an implicit bias toward temporal continuity when several solutions fit the data equally well. We thank the reviewer for raising this aspect and we have now added this clarification to the paper (App. A).

With respect to the robustness of the lags themselves, the lag matrix SS is shared across all time points, which acts as a natural regularizer, in addition to the explicit penalties we apply across its rows and columns (Eq. 4, Sec. 3.2). Therefore, the whole recording constrains which lags each circuit uses even when the coefficients vary over time.

Regarding the reviewer's question of whether the fitting procedure estimates similar lag patterns across initialization seeds and reasonable choices of hyperparameters, we would like to refer the reviewer to the stability experiments in the appendix. In particular, we found that across 100 seeds on the first synthetic dataset ML-dLDS reached a circuit correlation of 0.96 ± 0.08 and a lag support of 97% ± 10%, and across 200 seeds on the second synthetic dataset it reached 0.92 ± 0.10 and 76% ± 26% (App. D.2, Fig. 9).

In a different experiment, we ran ML-dLDS and the baselines across 200 seeds and found that ML-dLDS significantly outperformed the baselines on circuit correlation (Table 2). In addition, in a separate experiment, we ran an ablation variant of ML-dLDS without a lag penalty in one form, and in a different form kept only its ℓ2\ell_2 term, i.e., not the complete penalty. We found that in those cases the lag support dropped sharply (Fig. 3G), which supports the structured lag penalty and each of its components.

Finally, when we swept the penalty weights λ1\lambda_1 and λ3\lambda_3 over a grid, we found that lag-support recovery stays stable and that the main-text setting sits inside a broad 100% plateau (Fig. 8). We observed the same stability across the lag-weight profile and the maximum allowed lag (Fig. 7, Fig. 6).

W2. Lag recovery on data beyond the model, including data with no lag structure.

We thank the reviewer for this important question and agree that the robustness of ML-dLDS to spurious lags should be tested. We kindly note that the original submission partially addressed this question in the appendix. In particular, when we fit the second synthetic dataset with a maximum lag larger than the true one ML-dLDS recovered the lag structure correctly and did not assign dynamics to the extra lag (App. D.1, Fig. 6A to C).

Following the reviewer's suggestion we now generated two additional synthetic datasets that test this point: 1) data in which every circuit acts at lag one only and 2) data generated from a vector autoregression (VAR) with a separate matrix per lag.

For the first dataset we generated three circuits that act at lag one only, using the same generator as in our first synthetic dataset, and fit ML-dLDS with a maximum of three lags, i.e., leaving room for spurious lags. Across five seeds ML-dLDS recovered lag one as the dominant lag of every circuit and reconstructed the data with a relative squared error of 0.05 for the learned circuits (Fig. 15).

In the second additional experiment we designed VAR-driven data in which each circuit is associated with a single distinct lag, xt=∑δ=1ΔAδxt−δx_t = \sum_{\delta=1}^{\Delta} A_\delta x_{t-\delta}. We then fit ML-dLDS to this data across five random seeds and found that it recovered the dominant lag of every circuit, with a correlation of 0.98 between the true and recovered lag profiles (Fig. 17).

  • We hope these clarifications and the new experiments address the reviewer’s concerns and questions.
Reviewer_583XOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Multi-Lag decomposed Linear Dynamical Systems (ML-dLDS), which represents time-varying dynamics using shared linear interaction operators, sparse circuit-specific lag profiles, and time-varying circuit-lag activations. The model jointly estimates latent states and dynamical components through alternating optimization with sparsity regularization. Validation includes two synthetic datasets with ground-truth recovery, robustness analyses and ablations, one multi-region mouse recording session, and COVID-19 case data.

Critical strengths
  1. Explicitly modeling concurrent interactions at multiple lags addresses a relevant limitation of single-lag dLDS and supports analyses of interactions operating at different delays.
  2. The model explicitly distinguishes the lags available to each circuit from its momentary activation, separating persistent temporal structure from changing network engagement. This provides a useful framework for formulating testable hypotheses about how neural interactions are recruited over time.
Critical weaknesses
  1. The orthogonality penalty favors non-overlapping lag profiles, but its effect on recovering circuits that genuinely share delays is not established.
  2. The claimed parameter-efficiency gain is confounded by comparing a low-dimensional latent model with an observed-space VAR. Figure 4 establishes the behavioral relevance of the latent representation, but does not independently validate the physiological interpretation of individual lag-specific circuits.
Questions for authors
  1. How robust is lag recovery when the generating dynamics are nonlinear, and what measurable interpretability or data-efficiency benefits does the linear operator dictionary provide relative to a matched nonlinear delayed model?
  2. What biological insight do the lag-specific circuits in Figure 4 provide beyond improved reconstruction and choice decoding? Can the authors validate specific pathway- or behavior-dependent interactions and clarify whether the signs of the inferred operators support any excitation/inhibition interpretation (Similar to the E/I interpretation of Netformer)?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 583X

We thank the reviewer for the thorough review and for recognizing that modeling concurrent interactions across multiple lags addresses a relevant limitation of single-lag dLDS and that separating each circuit's available lags from its momentary activation supports testable hypotheses.

W1. Orthogonality and shared delays.

We wanted to clarify that our motivation for the orthogonality penalty (Eq. 4) is threefold: 1) interpretability of the multi-scale dynamical process, 2) robustness of the fitting procedure, and 3) a biologically motivated assumption that distinct processes may occur at different timescales.

For the first, we note that if every circuit acted at every lag, the decomposition could not reveal which interaction operates at which timescale and would create an everything-is-everywhere view that is hard to interpret. In particular, the orthogonality penalty helps with this by promoting the distribution of different circuits to different lags (but does not force it), such that each circuit can be read as a distinct process at its own timescales.

For the second, the penalty further constrains the sparsity structure and limits the number of available solutions, which helps in large systems where the solution is naturally ambiguous.

For the third, we focus on the biological motivation found in previous literature (Semedo et al., 2022, feedforward and feedback interactions between visual areas at different lags; Tauste Campo et al., 2015, interaction delays that differ by pathway across cortical areas) which often suggests that communication within the brain operates at multiple timescales.

We also wanted to emphasize that this possibility of genuinely shared lags is the exact reason we apply this assumption as a soft penalty added to the overall cost instead of a hard constraint. This penalty pushes the model to balance this regularization against data fidelity, which means that if the data genuinely contain circuits that share lags, the model will prioritize such a solution in order to minimize the data loss. Notably, this choice to apply it as a penalty also allows the weight of the penalty to be tuned according to the strength of the prior one wishes to impose (or set it to zero to remove the penalty).

To test this directly, we now generated an additional synthetic experiment in which two circuits act at the same lag, and we fit the lag gains with and without the orthogonality penalty while holding the circuits fixed. We found that ML-dLDS kept both circuits on the shared lag with full lag support in both cases (Fig. 16, App. L.2).

W2. Parameter efficiency confound.

We thank the reviewer for pointing out the confound when comparing to the full VAR. We agree that this is an important point that should be clarified in the paper and we have now clarified it (Sec. 4.3).

A VAR can be fit in two ways, either in the full observed space or in the latent space, and we found that ML-dLDS has an advantage in both. In particular, in the observed space, as detailed in the original paper, the VAR indeed requires far more parameters to match ML-dLDS. For the latent space, we now ran an additional experiment on both synthetic datasets, where we compared ML-dLDS against a VAR fit with the same number of latent dimensions (P=6P = 6). We found that the two models use a comparable number of parameters (117 for ML-dLDS against 108 for the matched VAR on the first dataset, and 120 against 144 on the second), but the VAR reaches a much higher reconstruction error (0.449 against 0.0002 for ML-dLDS on the first dataset, and 0.239 against 0.019 on the second) (Fig. 18, Table 5).

W3 and Q2. Physiological validation of the lag-specific circuits and an excitation/inhibition reading.

We thank the reviewer for raising this point, and we agree that validating the physiological interpretation of the individual circuits is important.

To address this, we now added a new experiment with an analysis of the signs of the inferred circuits (Fig. 20). In particular, we mapped each latent circuit to the neuron space through the loading matrix and read the signs of the effective area-to-area interactions at lag one and lag four. We found that the interactions are mostly excitatory, with an inhibitory component that is concentrated across areas and grows at longer lags (Fig. 20, App. O).

AuthorsOpenReview ↗

Response to Reviewer 583X, part b

Q1. Lag recovery under nonlinear dynamics and benefits over a matched nonlinear model.

We thank the reviewer for raising the question and we address its three parts in turn.

First, regarding robustness to nonlinear dynamics, following the reviewer's suggestion we now added an experiment in which we generated data from a mildly nonlinear delayed system and fit ML-dLDS to it. We found that ML-dLDS recovered the dominant lag of every circuit, with a lag-profile correlation of 0.98 (Fig. 17B, App. L.3).

Second, regarding interpretability, linear dynamical operators provide us with the advantage that each entry, e.g., fk[i,j]f_k[i,j], can be directly interpreted as the effect of latent jj on latent ii at the lags where circuit kk acts. Notably, this linear representation also benefits from the standard tools of LDS analysis, e.g., characterizing each circuit by its eigenvalues. A complex nonlinear map, in contrast, does not necessarily admit this reading and largely cannot be summarized by such properties.

We note that in our model, although each circuit is linear, the small set of such circuits, combined through the time-varying coefficients across multiple lags, can capture an expressive, non-stationary, and overall nonlinear signal shaped by the dynamics.

To test this directly, we now added a new experiment comparing ML-dLDS to a matched nonlinear model (a perceptron fit on the same nonlinear data, App. N). We found that the perceptron requires many more parameters (1349 weights versus about 120) to reach a comparable reconstruction error (0.41 versus 0.53) and does not provide us with the advantage of recovering the compositional, interpretable units that support the dynamics (Fig. 19, App. N).

Third, regarding data efficiency, we now added a new experiment in which we fit both models on training segments of increasing length of the same nonlinear data and scored them on a held-out segment. We found that ML-dLDS reaches a lower held-out error than the perceptron at every training length, with an error that stays flat across lengths, whereas the perceptron approaches it only at the longest segment (Fig. 21, App. P).

We hope these clarifications and the new experiments address the reviewer's concerns.

Thank you for the responses and additional experiments, which address my major concerns. I am happy to raise my score.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Author summary for the AC and reviewers: additional experiments and clarifications

We thank the reviewers for their engagement and constructive feedback as well as the AC for handling our submission. We summarize below the key additional experiments and clarifications we added during the rebuttal, which we hope address the reviewers' concerns. The additions are marked in pink in the revised paper we uploaded to OpenReview.

Additional experiments

  1. To examine the robustness of ML-dLDS to cases where all circuits act at a lag of one step, we added a new experiment in which every circuit acts at a lag of one step only and fit ML-dLDS with up to three lags allowed. We found that ML-dLDS assigned the dominant lag of every circuit to one step, with a relative squared reconstruction error of 0.05 across five seeds (Fig. 15, App. L.1; Reviewers p4sv, Dfp8).

  2. To check whether the orthogonality penalty of ML-dLDS obscures cases where two or more circuits genuinely share a lag, we added a new experiment in which two circuits act at the same lag and fit the lag gains with and without the penalty. We found that ML-dLDS kept both circuits on the shared lag with full lag support in both cases (Fig. 16, App. L.2; Reviewer 583X).

  3. To assess whether ML-dLDS recovers the correct lags on data generated by other models, we added two new experiments, one with data generated by a vector autoregression with a separate matrix per lag and one with data generated by a nonlinear delayed system. We found that ML-dLDS recovered the dominant lag of every circuit on both, with a lag-profile correlation of 0.976 across five seeds (Fig. 17, App. L.3; Reviewers 583X, p4sv, Dfp8).

  4. To remove the confound of comparing ML-dLDS against a VAR fit in the observed space, we added a new experiment in which we compared ML-dLDS against a VAR fit in the same latent dimension. We found that the two models use a comparable number of parameters, 117 against 108 and 120 against 144, and that ML-dLDS reaches a far lower reconstruction error, 0.0002 against 0.449 and 0.019 against 0.239 (Fig. 18, Table 5, App. M; Reviewers 583X, Dfp8).

  5. To evaluate what the linear circuits of ML-dLDS offer against a nonlinear model, we added a new experiment in which we compared ML-dLDS against a perceptron fit on the stacked lagged states of the nonlinear data. We found that ML-dLDS needs far fewer parameters than the perceptron (117 vs 1349) to reach a comparable in-sample error (0.53 vs 0.41), and that it returns three shared circuits with their per-lag gains whereas the perceptron exposes no circuits and no lag profile (Fig. 19, App. N; Reviewers 583X, Dfp8).

  6. To compare the data efficiency of ML-dLDS versus the same nonlinear model, we added a new experiment in which we fit both models on training segments of 200 to 600 time points and scored them on a held-out segment. We found that ML-dLDS reaches a lower held-out error compared to the perceptron at every training length, and that the perceptron approaches it only at 600 points (Fig. 21, App. P; Reviewers 583X, Dfp8).

  7. To test whether the signs of the inferred circuits support an excitation and inhibition reading, we added a new analysis of the mouse recording in which we mapped each circuit to neuron space and read the effective area-to-area interactions. We found that the interactions are excitatory dominant with an inhibitory component concentrated across areas that grows at longer lags (Fig. 20, App. O; Reviewer 583X).

Clarifications

  1. We clarified that ML-dLDS addresses a problem dLDS cannot, in particular capturing both the circuits and their lags, which dLDS would collapse onto a single lag. We also showed that a single-lag fit mixes the true circuit with the state's own autocorrelation, which also distorts the recovered circuits themselves (i.e., not only their lags). We further clarified that a naive multi-lag extension of dLDS would admit many equivalent solutions when past states at different lags carry similar information, which means that this problem requires a new formulation that calls for a new model. For that, ML-dLDS includes 1) a structured lag penalty, 2) a modified inference procedure, and 3) a new initialization, and together these address this ambiguity (Section 3.2, App. J; Reviewer Dfp8).
AuthorsOpenReview ↗

Author summary for the AC and reviewers: additional experiments and clarifications - PART B

  1. We emphasized that the recovered lags are stable across seeds and hyperparameters. Across 100 seeds on the first synthetic dataset, ML-dLDS reached a circuit correlation of 0.96 ± 0.08 and a lag support of 97% ± 10%, and the lag-support recovery sits inside a broad plateau when we sweep the penalty weights over a grid (App. D.2, Fig. 9, Fig. 8; Reviewer p4sv).

  2. We highlighted that the structured lag penalty is critical for lag recovery. When we removed it in an ablation experiment, or kept only its L2 term (i.e., not the full penalty), the lag-support recovery dropped sharply (Fig. 3G, Table 2). In addition, single-lag dLDS reached a much lower circuit correlation than ML-dLDS (Table 1; Reviewer Dfp8).

  3. We noted that ML-dLDS does not assign dynamics to lags that the data do not support. When we fit the second synthetic dataset with a maximum lag larger than the true one, ML-dLDS recovered the lag structure correctly and left the extra lag empty (App. D.1, Fig. 6A to C; Reviewer p4sv).

  4. We explained that we chose baselines that produce components analogous to those of ML-dLDS, e.g., circuits and dynamics, which allows the comparison to show whether ML-dLDS recovers them more or less accurately. Nonlinear methods like LFADS provide uninterpretable latent dynamics and do not learn a set of circuits analogous to ML-dLDS. Methods like mDLAG focus on delayed shared latents across areas and do not learn a dynamical operator, which makes a direct comparison of the recovered components ill-defined (Reviewer Dfp8).

We hope this consolidated summary of the key additional experiments and clarifications is useful and that we addressed the reviewers' concerns.

We would be grateful if the reviewers and the AC could consider our rebuttal during the reviewer-AC discussion period.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.